Malware signature builder and detection for executable code

Abstract

 A method for implementing malware signature builder and detection for executable codes, according to which the op-codes of the executables are inspected, while disregarding other parameters of the executables-Signatures are generated from common engines of executable codes and a large number of malwares of the same family are represented by a small number of signatures. Then known and unknown malwares are identified using the small number of signatures.

Description

Field of the Invention

[0001] The present invention relates to the field of malware detection. More particularly, the invention relates to malware signature generation and detection.

Background of the Invention

[0002] Detecting a malicious software (hereinafter, "malware"- which is any program or file that is harmful to a computer) by a signature is a common practice in any software security tool. By generating a signature which uniquely identifiers a malware, one may prevent the propagation of this malware through networks and stop it from spreading to other computers.

[0003] A signature of an executable (a file that contains a program and that is capable of being executed or run as a program in the computer) may be generated in any number of techniques. For example, one may use an inherent binary sequence in the executable and hash this sequence using a cryptographic hash function such as MD5 (an algorithm that is used to verify data integrity through the creation of a 128-bit message digest from unique data input as a fingerprint is to the specific individual) or SHA1 (a cryptographic hash function). The generated hash can be used as a signature. Another common technique is to incorporate details which reside in the executable headers. For example, the libraries it is linked to.

[0004] The generated signature must be as reliable as possible. The false positive and false negative rate must equal zero (or get as close as possible to zero). To achieve this goal, signature generation is usually performed by experts, and involves rigorous inspection of the suspected code. Only after identifying a common code (a code that is generated by compilers, linked libraries etc.) the expert can extract the pieces of code which uniquely identify the malware itself.

[0005] One of the most obvious shortcomings of signature based detection is handling new, still in the wild, malware. A signature specifically tailored to a given executable is designed not to match other executables, which leaves an open door for newly generated malware.

[0006] Another obvious problem is the sheer amount of signatures one must maintain and check against, when inspecting suspected executables.

[0007] Current techniques for generating signatures are either manual (expert driven) or automatic. Experts rely on their ability to disassemble the suspected executable, analyze the flow of the code and identify code sections which are unique to the suspect executable. Automated systems cannot rely on flow analysis and resort to data mining (or similar) techniques. For example, the Polygraph (James Newsome, Brad Karp, Dawn Song, "Polygraph: Automatically Generating Signatures for Polymorphic Worms," S&P05, pp. 226-241, 2005) system tries to find matching patterns in a flow of suspected network traffic. Such common patterns are then converted to signatures.

[0008] It is therefore an object of the present invention to reduce the number of signatures used to detect malwares, thus speeding-up the malware-detection process.

[0009] It is another object of the present invention, to be able to detect new unknown malware.

[0010] Other objects and advantages of the invention will become apparent as the description proceeds.

Summary of the Invention

[0011] The present invention is directed to a method for implementing malware signature builder and detection for executable codes, according to which the op-codes of the executables are inspected, while disregarding other parameters of the executables. Signatures are generated from common engines of executable codes and a large number of malwares of the same family are represented by a small number of signatures. Then known and unknown malwares are identified using the small number of signatures.

[0012] In one embodiment, benign executables are disassembled by a disassembler and the op-code sequences of the benign executable are stored in a first DB and each malware executable is disassembled by the disassembler to a sequence of op-codes. All common op-codes substrings are found by a string comparator (e.g., using an LCS process) for each pair of malware executables. A set of common op-codes substrings between any two malware executables is composed by the string comparator, which also checks which of the op-code substrings appear in at least one benign executable op-code sequence stored in the first DB. The common op-code sub-strings are ignored and a set of common op-code substrings, which do not appear in any benign executable op-code for each pair of malware executables is identified by the string comparator. All sets of common op-code substrings are stored in a second DB, as candidate signatures.

[0013] The string comparator may be used to find, for each candidate signature, the malwares in which the candidate signature appears as well as the minimal set of candidate signatures identifying all the malware executables. Then the minimal set is stored in the second DB.

[0014] The op-code sequences may be converted into regular expressions, which can be matched using binary sequences.

[0015] A single file signature may be generated by finding, for each offset in the op-code string of the file, the shortest substring starting at this offset, which does not appear in any clean file.

[0016] The shortest substring starting at this offset, which does not appear in any clean file, may be found by receiving as an input, a malware executable and a benign set of files; disassembling each benign executable by a disassembler; storing the op-code sequences of the benign executable into the first DB disassembling the malware executable into an op-code sequence; and generating by the string comparator, all common op-code substrings between the op-code sequence of the malware and each op-code sequence f the benign files stored in the first DB. The exact offsets in the malwares' op-code sequence from which the common substring started are recorded by the string comparator, for each common substring generated. A possible signature being the shortest string starting from this offset which is not contained in a common op-code substring is denoted by the string comparator, for each offset in the malwares' op-code sequence. Then the set of possible signatures is stored in the second DB and the op-code sequences are converted into regular expressions, to be matched as binary sequences.

[0017] Signatures may be converted to regular expressions by extracting by the string converter, op-code substrings (signatures) from the second DB and converting the op-code substrings into regular expressions using a lookup table; reducing the length of the resulting regular expressions by string converter by adding "binary masks" as matchable elements; and storing the regular expressions in the second DB.

[0018] Signatures may be detected on executables by disassembling by the disassembler any examined executable file into a sequence of op-codes; detecting by the string comparator if one of the signature regular expressions stored in the second DB exists in the examined executable op-code sequence and issuing a corresponding "passed" or "failed" indication. Whenever the string comparator finds at least one signature regular expression in the examined executable op-code sequence, a "failed" indication is issued and the examined executable is classified as malware. Otherwise, a "passed" indication is issued and the examined executable is classified as safe.

[0019] Signature regular expressions may be checked by the string comparator for similarity by defining a similarity relation between strings and matching by the string comparator, a signature regular expression to similar strings in the examined executable file. Whenever a high similarity value is found, a "failed" indication is issued and the examined executable is classified as malware. Whenever a low similarity value is found, a "passed" indication is issued and the examined executable is classified as safe.

[0020] Executables may be inspected while being in-transit at the network layer by creating several versions from each extracted signature and by generating smaller signatures from each of the signatures stored in the second DB, using a sliding window technique.

[0021] Whenever the number of smaller signatures is limited, executables may be inspected by generating by the string converter, two small signature for the start and end of the examined executable; determining a reasonable minimum packet size and discarding each packet breaching the minimum. Whenever at least one of the two signature regular expressions are found in the examined executable op-code sequence, a "failed" indication is issued and the examined executable is classified as malware. Whenever none of the two signature regular expression is found in the examined executable op-code sequence, a "passed" indication is issued and the examined executable is classified as safe.

Brief Description of the Drawings

[0022] The above and other characteristics and advantages of the invention will be better understood through the following illustrative and non-limitative detailed description of preferred embodiments thereof, with reference to the appended drawings, wherein:

• Fig. 1 illustrates a system for reducing the number of signatures used to detect malwares, according to an embodiment of the invention;
• Fig. 2 is a flow diagram illustrating the generation of signatures for a family of malwares;
• Fig. 3 is a flow diagram form illustrating the generation of a small set of signatures for a family of malwares;
• Fig. 4 is a flow diagram form illustrating the generation of signature for a single malware; and
• Fig. 5 illustrates a detection process that is carried out on a suspected executable.

Detailed Description of Preferred Embodiments

[0023] The present invention discloses a method and use for signature extraction from malwares. Accordingly, families of malwares, such that two members of the same family that share a common "engine", are identified.

[0024] It is well known that there are malware generation utilities which use a common engine to create new malware. The preset invention proposes to generate signatures from said engine, thus allowing a large number of malwares of the same family and even still unknown members of the same family to be identified by a small number of signatures. Using a small number of signatures to represent a large number of malwares alleviates the malware detection process, making it more efficient both in terms of resource allocated, CPU usage and time complexity. Moreover, future malware designers, that will use the same shared engine, will also be detectable by the proposed generated signatures.

[0025] According to the present invention, only the portion of a machine language instruction that specifies the operation to be performed (hereinafter "op-codes") of an executable are inspected, while disregarding any parameters given. This is a new technique, which has not been considered before. The reasoning behind it is that one must be aware that malware designers will strive to hide such common engines using a broad range of techniques. For example, these common engines may be located in different locations inside different executables, they may be mapped to different addresses in memory or even perturbed slightly. By using the op-codes, methods used by malware designers can be overcome.

Multiple Files Signature Generation

[0026] A system for reducing the number of signatures used to detect malwares according to an embodiment of the invention, is shown in Fig. 1. The system comprises a computer 100, connected to a first database 10 (DB 10) and a second database 20 (DB 20). Computer 100 includes a disassembler (used to convert a program in its executable, ready-to-run, form or object code into a representation in some form of assembler language so that it is readable) 200, a string comparator 300 and a string converter 400.

[0027] Fig. 2 illustrates in a flow diagram of the process which is performed according to an embodiment of the present invention. Accordingly, generating signatures starts at a preprocessing step. In step 21, a DB 10 that holds samples of benign executables is composed by disassembling each benign executable by disassembler 200, and storing the op-code sequences of said benign executable in DB 10. According to one embodiment of the present invention the IDA-Pro (DataRescue, Liege-Belgium) disassembler is used. A large and diverse database is of the essence, as the size and the variety of the database directly affect the false-positive rate (i.e., identifying an executable as malware, even though it is perfectly safe). According to one embodiment of the present invention, a database of 23,000 benign executables, ranging over most common operating systems is established.

[0028] Step 22 illustrates that given a family of malwares (signature generation for stand-alone malware will be discussed later on), each malware executable is disassembled by disassembler 200 into a sequence of op-codes. According to one embodiment of the present invention the IDA-Pro disassembler is used. Step 23 illustrates that for each pair of malware executables, string comparator 300 finds all common op-codes substrings, using any linear time LCS (Largest Common Substring) algorithm. Step 23 further illustrates that after string comparator 300 composes the set of common op-codes substrings between any two malware executables, it checks which of said op-codes substrings appear in at least one benign executable op-code sequence in DB 10, and these common op-code sub-strings are ignored. The end result of this step is that for each pair of malware executables, there exists a set of common op-code substrings, which do not appear in any benign executable op-code. These sets of common op-code substrings are stored by string comparator 300 in DB 20. Each of these op-code substrings, which are the candidate signatures, can be used as a signature.

[0029] According to one embodiment of the present invention, all of the candidate signatures stored in DB 20 are used.

[0030] According to another embodiment of the present invention, a small number of signatures are desired (e.g., in real-time filtering), and a small as possible number of the candidate signatures is used, as each op-code substring can potentially cover more than just one pair of malware executables. The signatures found using this process are likely to contain the shared code segments of the malware executables, which, in turn, define the common "engine" of said malwares executables. Moreover, such common engines are very likely to match future versions of these malware executables, as future versions normally employ said common engine. Referring to Fig. 3, steps 31-33 are identical to the step 21-23 specified in Fig. 2. However, in order to find a small number of signatures from the set of candidate signatures, as illustrated in step 34, string comparator 300 finds for each candidate signature the malwares it appears in (as it may appear in more than two). String comparator 300 then finds the minimal set of candidate signatures such that all the malware executables are identified by said set. This is done by converting the problem of the minimal set of candidate signature to the problem of coloring of the malware files set.

[0031] Each candidate signature is converted to a color, and as such, colors all the malware executables it appears in. The combinatory problem to find a minimal set of colors such that all malware executables are colored at least once, is then solved by string comparator 300. As this problem is NP-complete (a problem is called NP (Nondeterministic Polynomial) if its solution can be guessed and verified in polynomial time. If a problem is NP and all other NP problems are polynomial-time reducible to it, the problem is NP-complete), according to one embodiment of the present invention, a simple greedy algorithm is used, which yields a good approximation of signatures to be used. For example, in a family of 80 malwares examined, 15 signatures were enough to cover the whole family. The found set of candidate signatures is then stored by string comparator 300 into DB 20.

[0032] The result of this stage is a set of signatures, which are composed of a sequence of op-codes substrings, as described above. According to one embodiment of the present invention, string converter 400, as will be detailed below, is then responsible to convert said op-code sequences into regular expressions, which can be matched using binary sequences.

Single File Signature Generation

[0033] In another embodiment of the present invention, a signature for a stand-alone file is generated (e.g., said file had no common substrings with other malware).

[0034] Generating a signature for a single malware has been extensively studied in the past. Any method which generates a signature given a set of input symbols (in our case, vla-codes) and a set of "clean" specimen can be used. According to the invention, given a malware executable, the system finds, for each offset in the op-code string of said file, the shortest substring starting at this offset, which does not appear in any clean file.

[0035] Fig 4 illustrates how to find the shortest substring starting at this offset, which does not appear in any clean file. Device 100 receives as input a malware executable and a benign set of files. In step 41, DB 10 is formed by disassembling each benign executable by disassembler 200, and storing the op-codes sequences of said benign executable into DB 10. According to one embodiment of the present invention, the IDA-Pro disassembler is used. Moreover, disassembler 200 disassembles said malware executable into an op-code sequence. In step 42, string comparator 300 generates all common op-code substrings between the op-code sequence of the malware and each op-code sequence of the benign files stored in DB 10. In step 43, for each common substring generated, string comparator 300 records the exact offsets in the malwares' op-code sequence from which the common substring started. In step 44, for each offset in the malwares' op-code sequence, string comparator 300 denotes the shortest string starting from this offset which is not contained in a common op-code substring, as a possible signature. In step 45, string comparator 300 stores the set of possible signatures into DB 20 for future use.

[0036] The result of this stage is a set of possible signatures, which are composed of a sequence of op-codes substrings, as described above. According to one embodiment of the present invention, string converter 400, as detailed below, is then responsible to convert said op-code sequences into regular expressions, which can be matched in binary sequences.

Converting Signatures to Regular Expressions

[0037] According to one embodiment of the present invention, string converter 400 extracts op-code substrings (signatures) from DB 20, and converts said op-code substrings into regular expressions. This is done in order to get a compact, dense, easy to use signature. As each op-code has a finite set of binary representations, the generation of a regular expression is facilitated by a simple lookup table. However, the resulting regular expressions are rather long as the set of possible binary representations can be quite large. For example, the "add" op-code has 9 different binary representations. Moreover, the length of the regular expressions is rather long, as the smallest element matchable by most regular expressions implementations is a byte, some op-codes are shorter than a byte and encode registers information in this same byte.

[0038] In an embodiment of the present invention, the length of the resulting regular expressions is reduced by string converter 400, using the following augmentation to regular expression operator. Namely, string converter 400 adds "binary masks" (a screen of numbers that tells which numbers to look at underneath. In a binary mask, a "1" over a number means "Look at the number underneath"; a "0" means "Don't look.") as matchable elements. For example, the following binary values are matched, "0xa1, 0xa2, 0xa3,..., 0xaf", using a single expression, "&0xaf", which implies that a byte X matches if and only if the bitwise (i.e., data at the bit level) and of X and 0xaf differs from 0. Hence, string converter 400 reduces the length of the generated regular expressions considerably. After creating said regular expressions from each of the signatures, string converter 400 stores said regular expressions into DB 20.

Signature Detection on Executables

[0039] As disclosed above, each generated signature stored in DB 20 is converted into a regular expression. Fig. 5 illustrates a detection process that is carried out on a suspected executable. In step 51, an examined executable file is disassembled into a sequence of op-codes by disassembler 200. According to one embodiment of the present invention, the IDA-Pro disassembler is used. In step 52, string comparator 300 then detects if one of the signature regular expressions stored in DB 20 exists in the examined executable op-code sequence and issues a corresponding "passed" or "failed" indication. This is achieved by employing a finite automaton of each of the regular expressions signatures on the executable examined op-code sequence, by string comparator 300. If string comparator 300 finds at least one signature regular expression in the examined executable op-code sequence, a "failed" indication is issued and the examined executable is classified as malware, as illustrated in step 53. If however, no signature regular expression is found in the examined executable op-code sequence, a "passed" indication is issued and the examined executable is classified as safe, as illustrated in step 54.

[0040] In an embodiment of the present invention, No Operation Performed (NOP- a command that effectively does nothing) op-codes are ignored by string comparator 300 when matching signature regular expressions in said examined executable op-code sequence, to prevent the simple obfuscation of malware by inserting NOPs in the instruction sequence.

[0041] In another embodiment of the present invention, the signature regular expressions are checked by string comparator 300 for similarity, and not only for exact match in the examined executable op-code sequence. Namely, a similarity relation between strings is defined, and a signature regular expression is matched by string comparator 300 to similar strings in the examined executable file. When a high similarity value is found, a "failed" indication is issued by string comparator 300 and the examined executable is classified as malware. If however, a low similarity value is found in the examined executable op-code sequence, a "passed" indication is issued and the examined executable is classified as safe.

[0042] According to another embodiment of the present invention, inspection is made to executables that are in-transit at the network layer. One may not always be able to capture the entire length of an executable, as the executable may be divided in parts and each part is routed differently (when using, for example, the TCP/IP protocol).

[0043] To overcome the aforementioned problem, string converter 400 creates several versions from each extracted signature. Each version is a smaller signature, covering a possible portion of the original signature. Using this technique the probability of identifying a packet in transit as a possible part of malware executable is boosted.

[0044] String converter 400 generates from each of the signatures stored in DB 20 smaller signatures using a sliding window technique. If the lower bound on the portion of the executable the system can inspect is K (measured in bytes, for each consecutive K op-codes of the executable under examination, string converter 400 generates a signature. Each of the K consecutive op-codes is compared by string comparator 300 to the smaller signatures extracted by string converter 400. If string comparator 300 finds at least one signature regular expression in the examined executable op-code sequence, a "failed" indication is issued and the examined executable is classified as malware. If however, no signature regular expression is found in the examined executable op-code sequence, a "passed" indication is issued and the examined executable is classified as safe.

[0045] According to another embodiment of the invention, the number of smaller signatures is limited, and string converter 400 generates two small signature for the start and end of the examined executable, as the start and end of an executable can be ensured to be contained in a single packet, by limiting the size of the smallest packet (such truncated signatures also need to be validated against the benign file DB 10, to ensure there are no false positives). According to one embodiment of the invention, a reasonable minimum packet size is determined, and each packet breaching this minimum is discarded (this, for example, prevents an attacker from sending an executable byte by byte). If string comparator 300 finds at least one of said two signature regular expressions in the examined executable op-code sequence, a "failed" indication is issued and the examined executable is classified as malware. If however, none of said two signature regular expression is found in the examined executable op-code sequence, a "passed" indication is issued and the examined executable is classified as safe.

Exemplary Results

[0046] A family of 99 worms was studied. A total of 15 signatures were enough to cover 79 worms of the family. The other 20 worms did not have a common substring with other worms and needed a signature of their own.

[0047] The above examples and description have of course been provided only for the purpose of illustration, and are not intended to limit the invention in any way. As will be appreciated by the skilled person, the invention can be carried out in a great variety of ways, employing more than one technique from those described above, all without exceeding the scope of the invention.

Claims

1. Method for implementing malware signature builder and detection for executable codes, comprising:

a) inspecting the op-codes of said executables, while disregarding other parameters of said executables;

b) generating signatures from common engines of executable codes;

c) representing a large number of malwares of the same family by a small number of signatures; and

d) identifying known and unknown malwares using said small number of signatures.

2. Method according to claim 1, comprising:

a) disassembling benign executables by a disassembler;

b) storing the op-code sequences of said benign executable in a first DB.

c) disassembling each malware executable by said disassembler to a sequence of op-codes;

d) for each pair of malware executables, finding all common op-codes substrings by a string comparator;

e) composing a set of common op-codes substrings between any two malware executables, by said string comparator;

f) checking by said string comparator, which of said op-code substrings appear in at least one benign executable op-code sequence stored in said first DB;

g) ignoring said common op-code sub-strings;

h) for each pair of malware executables, identifying by said string comparator, a set of common op-code substrings, which do not appear in any benign executable op-code; and

i) storing by said string comparator, all sets of common op-code substrings in a second DB, as candidate signatures.

3. Method according to claim 2, wherein all common op-codes substrings are found by the string comparator using an LCS process.

4. Method according to claim 2, further comprising:

a) finding, for each candidate signature, by the string comparator, the malwares in which said candidate signature appears;

b) finding, by the string comparator, the minimal set of candidate signatures identifying all the malware executables; and

c) storing said minimal set in the second DB.

5. Method according to claim 4, further comprising converting the op-code sequences into regular expressions, which can be matched using binary sequences.

6. Method according to claim 1, wherein a single file signature is generated by finding, for each offset in the op-code string of said file, the shortest substring starting at this offset, which does not appear in any clean file.

7. Method according to claim 6, wherein the shortest substring starting at this offset, which does not appear in any clean file, is found by:

a) receiving as an input, a malware executable and a benign set of files;

b) disassembling each benign executable by a disassembler;

c) storing the op-code sequences of said benign executable into the first DB disassembling said malware executable into an op-code sequence;

d) generating by the string comparator, all common op-code substrings between the op-code sequence of the malware and each op-code sequence of the benign files stored in said first DB;

e) for each common substring generated, recording, by said string comparator, the exact offsets in the malwares' op-code sequence from which the common substring started;

f) for each offset in the malwares' op-code sequence, denoting, by said string comparator, a possible signature being the shortest string starting from this offset which is not contained in a common op-code substring;

g) storing the set of possible signatures in the second DB; and

h) converting said op-code sequences into regular expressions, to be matched as binary sequences.

8. Method according to claim 7, wherein signatures are converted to regular expressions by:

a) extracting by the string converter, op-code substrings (signatures) from the second DB and converting said op-code substrings into regular expressions using a lookup table;

b) reducing the length of the resulting regular expressions by string converter by adding "binary masks" as matchable elements; and

c) storing said regular expressions in the second DB.

9. Method according to claim 1, wherein signatures are detected on executables by:

a) disassembling by the disassembler any examined executable file into a sequence of op-codes;

b) detecting by the string comparator if one of the signature regular expressions stored in the second DB exists in the examined executable op-code sequence and issuing a corresponding "passed" or "failed" indication;

c) whenever the string comparator finds at least one signature regular expression in the examined executable op-code sequence, issuing a "failed" indication and classifying said examined executable as malware; otherwise,

d) issuing a "passed" indication and classifying said examined executable as safe.

10. Method according to claim 8, wherein signature regular expressions are checked by the string comparator for similarity by:

a) defining a similarity relation between strings;

b) matching by the string comparator, a signature regular expression to similar strings in the examined executable file;

c) whenever a high similarity value is found, issuing a "failed" indication and classifying the examined executable as malware; and

d) whenever a low similarity value is found, issuing a "passed" indication and classifying the examined executable as safe.

11. Method according to claim 10, wherein executables are inspected while being in-transit at the network layer by:

a) creating several versions from each extracted signature; and

b) from each of the signatures stored in the second DB, generating smaller signatures using a sliding window technique.

12. Method according to claim 10, wherein whenever the number of smaller signatures is limited, executables are inspected by:

a) generating by the string converter, two small signature for the start and end of the examined executable;

b) determining a reasonable minimum packet size;

c) discarding each packet breaching said minimum;

d) whenever at least one of said two signature regular expressions are found in the examined executable op-code sequence, issuing a "failed" indication and classifying the examined executable as malware; and

e) whenever none of said two signature regular expression are found in the examined executable op-code sequence, issuing a "passed" indication and classifying the examined executable as safe.

Drawing